Feature/kmeans++ #120
Merged
Feature/kmeans++ #120
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
marcocuturi
reviewed
Aug 10, 2022
marcocuturi
reviewed
Aug 10, 2022
tests/tools/k_means_test.py
Outdated
| x, k=k, weights=w, min_iterations=10, max_iterations=10, key=key1 | ||
| ).error | ||
|
|
||
| k, eps = 4, 1e-12 |
There was a problem hiding this comment.
maybe a bit too small? if this works with 1e-12 why not, but this sounds like a scale where numerical errors could easily pop up. Maybe more reasonable to use something like 1e-3 ?
There was a problem hiding this comment.
Changed eps=1e-3 and rtol=atol=1e-3. If using 64 bits, it passes with rtol=atol=1e-8, but haven't added a test for this, as it already takes some time.
marcocuturi
approved these changes
Aug 10, 2022
michalk8
added a commit
that referenced
this pull request
Jun 27, 2024
* Add initial impl. from `CR.Sparse` * Rename file, add to __init__ * Initial fixed point iteration * Add random initialization * Better KMeansState * Fix `cond_fn` * First working version * Clean output, use `tree_map` * Remove reference impl. and dead code * Add TODO * Expose `cost_rank` in `PointCloud` * Add initial kmeans++ implementation * Fix indexing bug * Remove `set` methods * Add tolerance to convergence check * Rename function * Store inner errors * Unify kmeans initializer interface, allow custom * Reorder arguments * Add strict convergence criterion * Add convergence iteration to output * Simplify `cond_fn`, use `max_iter - 1` * Clip cosine distance to `[0, 2]` * Require sqEucl geometry, allow arrays * Fix random/kmeans++ init * Remove normalization comment * Fix dividing by 0 when using weights * Add TODOs * Fix k-means++ init centroid * Use `jax.tree_util.tree_map` * Rename arguments, use sum instead of mean * Switch order * Fix `unique_indices=True` in segment sum * Don't compute assignment in `init_fn` * Fix weighting * Use center shift as convergence criterion * Remove old TODOs * Improve final assignment * Fix centroid/weight adjustment * [ci skip] Allow geometry with cosine cost * Fix cosine conversion, add test * Add more cosine -> sqeucl conversion tests * Add documentation * Add skeleton tests * Add kmeans++ tests * Add k-means initialization test * Fix bug when removing empty centroids * Finish tests * Increase tolerance * Address comments * Use smaller eps
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
In this PR:
k-means, includingk-means++initialization and weightingtree_utilwarnings infixed_point_loopDeviceArrayfor scaling of cost matricesrelated issue #111